in config.py, did min_sizes and max_sizes mean scale?

[squirrel16]

Nice work, thanks very much. But I have a little question:

    'min_sizes' : [30, 60, 111, 162, 213, 264],

    'max_sizes' : [60, 111, 162, 213, 264, 315],

Did this mean the scale of default boxes in ssd? Why did you set in this way?why is it different with 0.2-0.95 in the original caffe implementation?

[AndrewZhao]

the size of default box in caffe is also different from paper. min_size means the minimum object size in the current layer. you can change the sizes in order to fit your object size.

[jxlijunhao]

Actually, the paper and the origin caffe implement have some differences, so if you want to match this, you need read the origin caffe code

[chi0tzp]

@AndrewZhao could you elaborate a bit on that? What does the "minimum object size in the current layer" means? For instance, the last layer, which is 1x1, is for finding big objects, that would occupy the whole input image. What do the min_size=264 and max_size=315 mean in such a case? And also in the first layer, which is 38x38, what do the min_size=30 and max_size=60 mean?

Who should I set them for a new dataset? Should I measure the objects in their original dimensions and compute some statistics?

Many thanks.

[WeihongM]

@chi0tzp This size is relative to origin image size(such as 300x300). A helpful tutorial to understand this repo. Learning Note Single Shot MultiBox Detector with Pytorch

[jamiechoi1995]

see https://github.com/weiliu89/caffe/blob/ssd/examples/ssd/ssd_pascal.py#L299-L317